CF-CAM: Cluster Filter Class Activation Mapping

---

CF-CAM (Cluster Filter Class Activation Map) is an enhanced Class Activation Mapping (CAM) technique that improves the interpretability of deep neural networks by mitigating gradient perturbation and stabilizing feature maps. It extends Grad-CAM by introducing density-aware feature clustering (DBSCAN) and cluster-conditioned gradient filtering (bilateral filtering), resulting in more precise and robust heatmaps.

Key Features

• Density-Aware Channel Clustering: Uses DBSCAN to group semantically similar feature channels while eliminating noise-prone activations.
• Cluster-Conditioned Gradient Filtering: Applies bilateral filtering within clusters to refine gradient signals and enhance spatial coherence.
• Hierarchical Importance Weighting: Balances discriminative feature retention and noise suppression for faithful interpretability.

---

Installation & Setup

1. Install Dependencies

Ensure you have Python installed (>=3.8). Then, install the required libraries:

pip install -r requirements.txt

2. Download Pretrained Model Weights and Dataset

Download the fine-tuned ResNet50 model from Google Drive:

Pretrained Model Weights

Detailed information of the dataset:

Shenzhen Hospital X-Ray Set

Place the downloaded weight file in the root directory of this project. For the dataset, place the "CXR_png" folder in the root directory.

3. Prepare Input Images

Put test images inside the testimgs/ folder. Ensure that images are in common formats (e.g., .jpg, .png).

4. Run CF-CAM

Execute the following command to generate CAM heatmaps:

python CF-CAM.py

The generated heatmaps will be saved in the output/ directory.

---

CF-CAM Workflow

1. Model Initialization & Hook Registration

• Loads a pre-trained model (e.g., ResNet50) and selects a target convolutional layer (e.g., layer4[-1]).
• Registers forward and backward hooks to capture feature maps and gradients.

2. Image Preprocessing

• Loads an input image and resizes it to 224×224.
• Converts the image to a tensor and applies ImageNet normalization:

  mean = [0.485, 0.456, 0.406]
  std = [0.229, 0.224, 0.225]

3. Forward Pass & Class Prediction

• Runs the model on the input image.
• If no target class is specified, selects the top predicted class automatically.

4. Backward Pass

• Computes gradients of the target class with respect to the feature maps of the selected layer.

5. Feature Map Clustering (DBSCAN)

• Flattens feature maps into a (C, H*W) matrix.
• Computes the Euclidean distance matrix between feature channels.
• Uses adaptive DBSCAN clustering:
  • eps = 25th percentile of distance matrix
  • min_samples = max(2, 5% of total channels)
• Clusters semantically related channels and removes outliers (noise channels).

6. Gradient Filtering (Bilateral Filter)

• For each cluster:
  • Computes the mean gradient of the channels.
  • Applies bilateral filtering to preserve spatial coherence.
  • Uses filtered gradients to enhance interpretability.

7. Weight Calculation & CAM Generation

• Performs global average pooling on filtered gradients.
• Computes weighted sum of feature maps.
• Applies ReLU activation (torch.clamp(min=0)) and normalizes the heatmap.

8. Heatmap Visualization & Superimposition

• Upsamples the CAM heatmap to match the original image size.
• Converts CAM to a heatmap using OpenCV.
• Superimposes the heatmap onto the input image (40% heatmap + 60% original).
• Saves the final visualization to output/.